Lex Fridman PodcastRosalind Picard: Affective Computing, Emotion, Privacy, and Health | Lex Fridman Podcast #24
CHAPTERS
- 0:00 – 1:55
What affective computing really includes (beyond emotion recognition)
Lex asks how Rosalind Picard’s view of affective computing has evolved since she coined the term. Picard clarifies the original, broader definition: computing that relates to, arises from, or influences emotion—including emotion-like mechanisms inside machines, not just emotion detection.
- •Original definition: computing that relates to/derives from/influences human emotion
- •Not only recognizing emotion, but also responding intelligently to it
- •Also includes internal machine mechanisms that function like emotion
- •Human-computer interaction as a primary visible application area
- 1:55 – 5:02
Clippy as a lesson in emotionally unintelligent design
Picard uses Microsoft’s Clippy as an example of how systems can be ‘smart’ in task context yet socially and emotionally tone-deaf. The mismatch between user frustration and the assistant’s cheerful behavior illustrates why affect-sensitive interaction matters.
- •Clippy’s behavior amplified user frustration instead of adapting
- •Emotion mismatch is more damaging than simple task errors
- •Social-emotional interaction is harder than chess/Go-style problems
- •Early AI prioritized language/math/game tasks over human interaction
- 5:02 – 5:54
Have computer scientists gotten more empathetic? Diversity and the human factor
Lex probes whether computer science culture has improved in empathy over time. Picard argues the field is more diverse now, and that broader representation is essential for building technology that better reflects societal needs.
- •Computer science today includes a wider range of personality types
- •Diversity improves alignment with real societal needs
- •Brilliance isn’t limited to people who prefer computers over people
- •Risk if AI builders are disproportionately disconnected from humans
- 5:54 – 7:52
How hard is emotional intelligence for machines, really? Limits of narrow context
Picard explains that affective intelligence remains as hard as expected, and progress depends heavily on where society invests effort. She emphasizes core limitations: lack of consciousness, limited context understanding, and difficulty ‘reading between the lines.’
- •Difficulty prediction remains accurate; timelines depend on research focus
- •Systems still lack awareness/consciousness and implicit understanding
- •Success is likely in narrow, pre-specified contexts first
- •General emotional intelligence remains constrained by fundamental gaps
- 7:52 – 11:08
Putting on the brakes: surveillance misuse and the China scenario
The conversation shifts from technical difficulty to ethical urgency. Picard highlights the danger of emotion/affect sensing being used without consent—especially under authoritarian surveillance—where even subtle expressions could trigger punishment.
- •Greatest concern: deployment for coercive surveillance and control
- •Non-consensual affect sensing can punish thoughts/attitudes indirectly
- •Example: skeptical facial expressions while viewing political content
- •Affectiva’s stance: reject uses without prior informed consent
- 11:08 – 12:36
Deepfakes, physiological signal extraction, and the need to ‘jam’ sensing
Lex raises deepfakes as a strange form of protection (“it was fake”). Picard describes methods to extract heart rate/respiration from ordinary video—and the decision to also develop countermeasures to prevent misuse, redirecting research priorities toward safety.
- •Physiological signals can be inferred from video (heart rate, respiration)
- •Tampering can be detected—and also fabricated (spoofed signals)
- •Building ‘jamming’ tools trades off with advancing capability
- •Ethical worries shape what researchers choose to build next
- 12:36 – 22:22
Who benefits from AI? Inequality, incentives, and targeted regulation
Picard critiques AI being driven mainly by publication and profit, widening inequality. She supports targeted regulation focused on data ownership and restricting emotion recognition in sensitive contexts (e.g., hiring), extending protections similar to those around lie detection and medical data.
- •AI can concentrate wealth and widen social divides
- •Prefer ‘carrot’ incentives, but some regulation is necessary
- •People should own their data—not platforms
- •Emotion recognition should face restrictions similar to lie detectors
- •Mental-health-predictive signals deserve medical-grade protections
- 22:22 – 25:38
Should assistants read emotion? Suicide risk vs emotional manipulation for profit
Lex asks if Alexa/Siri-like systems should understand emotion. Picard notes real safety needs (e.g., distinguishing suicidal intent) but warns of monetization incentives: mood manipulation can change purchasing behavior, motivating a firewall between emotional access and sales systems.
- •Assistants already face suicidal-language inputs; tone/context matters
- •Two motivations: altruistic customer care vs profit optimization
- •Research shows mood affects spending (‘purse strings’ effects)
- •Call for separation between emotion-sensing agents and selling agents
- 25:38 – 30:32
Designing the ‘objective function’: helpful assistant vs button-pushing agent
They explore what an emotionally intelligent system should optimize—minimizing annoyance, maximizing happiness, or building resilience. Picard argues context matters (training self-control may require provocation), but her preference is AI that respectfully serves and extends human capability.
- •Objective depends on context: comfort vs resilience training
- •Provocation can be useful for learning self-regulation
- •Risk of ‘Brave New World’ style forced happiness/manipulation
- •Preferred model: respectful helper that empowers users
- 30:32 – 34:54
Expressed vs felt emotion: what cameras can infer (even from a ‘poker face’)
Lex asks about the gap between outward expression and inner feeling. Picard explains that faces alone are limited, but ordinary cameras can detect subtle color changes enabling inference of physiological activation (stress, breathing irregularities), especially when tracked over time.
- •Facial expressions can be masked, but physiology leaks through
- •Regular cameras can infer heart rate/respiration via subtle color shifts
- •What’s inferred is often arousal/activation—not nuanced feelings
- •Longitudinal patterns (constant surveillance) increase inference power
- •Thoughts and nuanced subjective feelings remain largely private
- 34:54 – 39:00
Best modalities and why wearables + phones can forecast tomorrow’s mood and stress
Picard discusses multi-modal sensing: wearables, smartphones, context, and weekly rhythms. She describes studies (notably with college students) showing strong forecasting of next-day stress/mood/health, with best results coming from combining signals—while noting wearables may offer more user control than cameras.
- •‘Best modality’ depends on what you want to know; everything is informative
- •Combining wearable + phone + context yields strongest prediction
- •Wearables alone still achieve high forecasting performance
- •Non-contact sensing can be scarier because it’s harder to notice/control
- •Control (easy opt-out) is central to reducing stress and enabling consent
- 39:00 – 44:31
From stress signals to seizure detection: Empatica Embrace, SUDEP, and deep brain mapping
Picard recounts how unusual skin conductance patterns (first noticed in autism research) led to recognizing seizure-related signatures and building the FDA-cleared Embrace device. She explains SUDEP (sudden unexpected death in epilepsy), why it often occurs when people are alone, and ongoing research linking peripheral signals to deep brain activity.
- •One-sided skin conductance anomaly revealed seizure-related brain activity
- •Seizures can be detected peripherally; generalized seizures show stronger signals
- •Embrace became FDA-cleared for seizure detection and alerting
- •SUDEP: major but underrecognized cause of life-years lost; often preventable
- •Research correlates skin conductance surges with dangerous post-seizure flattening
- •Implanted electrodes enable mapping deep regions tied to peripheral responses
- 44:31 – 48:31
Why FDA clearance is ‘agonizing’: safety, opacity, and innovation friction
Lex asks about the difficulty of getting computer-science-driven medical tech through the FDA. Picard praises the importance of safety testing but criticizes opaque, sometimes unexplained extra requirements that can slow life-improving technology.
- •FDA clearance is harder than publishing in top medical journals
- •Safety role is valuable and necessary
- •Frustration arises from lack of transparency and unexplained demands
- •Greater clarity could improve trust and reduce wasted effort
- 48:31 – 1:00:11
AI, embodiment, consciousness, and love—then a turn to faith, truth, and meaning
The discussion moves from practical AI to long-term possibilities: embodied robots, simulated consciousness, and whether humans could ‘fall in love’ with AI. Picard is skeptical that machine love can match healthy human relationships and closes by arguing against scientism—defending multiple ways of knowing (history, philosophy, love, faith) and emphasizing meaning beyond measurement.
- •AI can simulate attachment, but likely won’t match good human relationships
- •Embodiment increases engagement, attention, and compliance
- •Consciousness remains unsolved; appearance of consciousness is easy to fake
- •Rights for AI may become political regardless of true consciousness
- •Critique of scientism: science isn’t the only path to truth
- •Meaning, purpose, love, and faith address ‘why’ questions science can’t prove
